Google expands AI agent kit to five languages

Google has expanded its Agent Development Kit to support Go, TypeScript, Java and Kotlin alongside Python, giving the toolkit native coverage across five programming languages.

The product is aimed at software teams building AI agents in different runtime environments, including cloud services, web browsers, JVM-based applications and Android devices. All five versions share a common architecture and work with the same open protocols, according to Google.

Those protocols include the Model Context Protocol, which lets a tool exposed as an MCP server be called from any supported language, and Agent2Agent, which allows agents to communicate with remote agents over HTTP. That means developers can mix languages and frameworks within the same multi-agent system rather than standardising on a Python-only stack.

Google has also introduced a setup flow and project scaffolding through its command-line tools. Developers can generate working projects with A2A endpoints and deployment configurations for Cloud Run, Gemini Enterprise Agent Runtime or Google Kubernetes Engine.

Language spread

Google described Python as the reference implementation within ADK, with the widest range of integrations for services such as BigQuery, Spanner and Bigtable, as well as model adapters including Gemini, Anthropic, OpenAI and LiteLLM. It also highlighted Python's evaluation tooling for testing agent behaviour against datasets in continuous integration workflows.

Go is positioned for backend deployment, with support for typed input and output schemas through generics and a graph workflow runtime for deterministic control flow. Google said the Go version is suited to Cloud Run and GKE deployments because it compiles into a small static binary and is designed to handle large numbers of concurrent requests.

TypeScript support is built around Zod schemas rather than runtime reflection. Google said this lets developers define runtime input validation, JSON schema and editor type inference from the same source. The TypeScript edition also includes a browser bundle, allowing agents to run directly in a browser tab and use WebMCP and Chrome's on-device Prompt API.

For Java developers, ADK uses schema annotations and a fluent builder pattern, with event streaming handled through RxJava. The Java release also includes a Maven Dev UI plugin and integrations for Spring AI and LangChain4j, which Google said could make it easier to insert the toolkit into existing JVM software estates.

Kotlin support takes a different approach, using Kotlin Symbol Processing to generate tool schemas at build time rather than relying on runtime reflection. Google said this allows earlier error detection during builds and supports Android use cases, including local agent execution on-device, integration with LiteRT-LM and ML Kit, and links to Room, AppSearch and Android AppFunctions without requiring a network connection.

Interoperability focus

The broader strategy appears to be reducing the need for companies to create separate Python services purely to host agent workloads. Instead, teams can build agents in the language already used by the relevant application or platform while keeping interoperability through shared protocols.

That could matter for organisations with mixed software estates. Many companies run browser-based front ends, Go or Java backend services, and Android applications in parallel, making a single-language requirement difficult to enforce without adding engineering overhead.

Google's description of the toolkit also points to a shift in how agent software is deployed. Rather than limiting agents to server-side Python environments, the company is pushing the idea that they can live in web clients, mobile apps and highly concurrent cloud services, provided they can still exchange context and tool calls through common standards.

The release builds on a wider industry effort to make agent frameworks more modular and less dependent on one runtime. By backing MCP and A2A across all five language implementations, Google is presenting ADK as an interoperability layer as much as a developer toolkit.

In practice, a tool built once as an MCP server could be reused by agents written in Python, Go, TypeScript, Java or Kotlin, while agents in different languages could communicate with each other over A2A endpoints. For engineering teams trying to connect experimental AI projects with production systems, that may reduce some of the friction between prototyping and deployment.

Google said every version of ADK speaks MCP and A2A, so “each component of your system can use the language that fits it best.”